Increased diversity of peptidic natural products revealed by modification-tolerant database search of mass spectra

Abstract

Peptidic natural products (PNPs) include many antibiotics and other bioactive compounds. While the recent launch of the Global Natural Product Social (GNPS) molecular networking infrastructure is transforming PNP discovery into a high-throughput technology, PNP identification algorithms are needed to realize the potential of the GNPS project. GNPS relies on the assumption that each connected component of a molecular network (representing related metabolites) illuminates the “dark matter of metabolomics” as long as it contains a known metabolite present in a database. We reveal a surprising diversity of PNPs produced by related bacteria and show that, contrary to the “comparative metabolomics” assumption, two related bacteria are unlikely to produce identical PNPs (even though they are likely to produce similar PNPs). Since this observation undermines the utility of GNPS, we developed a PNP identification tool VarQuest that illuminates the connected components in a molecular network even if they do not contain known PNPs and only contain their variants. VarQuest revealed an order of magnitude more PNP variants than all previous PNP discovery efforts and demonstrated that GNPS already contains spectra from 41% of currently known PNP families. The enormous diversity of PNPs suggests that biosynthetic gene clusters in various microorganisms constantly evolve to generate a unique spectrum of PNP variants that differ from PNPs in other species.

After a decline in the pace of antibiotics discovery in the 1990s, antibiotics and other natural products are again in the center of attention as exemplified by the recent discovery of teixobactin¹. Previous studies of natural products mainly relied on low-throughput NMR-based technologies requiring large amounts of highly purified material that are often difficult to obtain. The key condition for enabling the renaissance of the antibiotics research is development of high-throughput computational discovery pipelines such as the recently launched Global Natural Products Social (GNPS) molecular network² that already contains over a billion of tandem mass spectra, a gold mine for bioactive compounds discovery. However, natural products identification algorithms are needed to transform antibiotics discovery into a high-throughput technology and to realize the promise of the GNPS project.

This study focuses on Peptidic Natural Products (PNPs), an important class of natural products with an unparalleled track record in pharmacology: many antibiotics, antiviral and antitumor agents, immunosuppressors are PNPs. PNPs are produced by Non-Ribosomal Peptide Synthetases (NRPS)³ and Ribosomally synthesized and Posttranslationally modified Peptides Synthases (RiPPS)⁴. NRPS and RiPPS synthesize Non-Ribosomal Peptides (NRPs) and Ribosomally synthesized and Posttranslationally modified Peptides (RiPPs), respectively. NRPs are not directly inscribed in genomes but instead are encoded by NRPSs using non-ribosomal code, with each A-domain in NRPS responsible for a single amino acid in NRP^{5, 6}. While RiPPs are encoded in the genome, the RiPP-encoding genes are often short, making it difficult to annotate them⁷.

PNP identification

Given a spectrum and a peptide database, peptide identification refers to finding a peptide in the database that generated the given spectrum. Identification of spectra derived from PNPs^8–11 is more difficult than traditional peptide identification in proteomics because many PNPs are non-linear peptides (e.g., cyclic or branch-cyclic) that contain non-standard amino acids and complex modifications.

Identification of non-linear peptides is only one of the two major challenges in PNP identification. In many cases, a PNP is absent in the database of known PNPs, but its modified or mutated variant is present in this database. Identification of an unknown PNP from its known variants is called variable identification (as opposed to standard identification when a PNP is present in the database). Similarly to the problem of variable identification of modified peptides in traditional proteomics^12–15, variable PNP identification is difficult because the computational space of this problem is several orders of magnitude larger than for standard PNP identification.

Since most PNPs form families of related peptides, variable identification is crucial for PNP discovery. Finding variants of known PNPs is important since these variants are sometimes more effective than the most abundant representatives of PNP families that currently dominate the PNP databases. The antifungal drug Cancidas¹⁶ or modified variants of vancomycin¹⁷ are some examples of variant PNPs that proved to be effective in clinical applications.

Spectral networks

Given a set of PNPs P₁, . . . , P_m, their peptide network is a graph with nodes P₁, . . . , P_m, and edges connecting two PNPs if they differ by a single modification or mutation (substitution, insertion or deletion)¹⁸. Each component in the peptide network defines a PNP family. In reality, we are not given PNPs P₁, . . . , P_m but only their spectra S₁, . . . , S_m. Nevertheless, one can approximate the peptide network by constructing the spectral network on nodes S₁, . . . , S_m, where spectra S_i and S_j are connected by an edge if they are similar, e.g., can be aligned against each other using the spectral alignment algorithm¹³.

Although spectral networks¹⁹ reveal spectra of related peptides without knowing what these peptides are, they have an important limitation—they work only when one of the spectra (nodes) in the connected component of the network corresponds to an unmodified peptide from a database (referred to as an unmodified parent). As the result, orphan components in the spectral network (components without annotated nodes) represent the “dark matter of PNPs” since the spectral network propagation approach^{18, 19} lacks ability to interpret them in the absence of an unmodified parent (Figure 1a).

Figure 1. Network-based and network-independent strategies for variable PNP identification.

Variant PNPs are colored by the same color as their known compounds in the database; modified/mutated amino acids are highlighted by darker color. (a) Network-based PNP identification starts from the standard identification of spectra (i) and construction of a spectral network (ii). Next, the network is annotated (iii) using the identified PNPs via the spectral network propagation approach. In this example, the network component on the left has a single unmodified parent colored green as the related PNP, while the component on the right is an orphan. Annotation propagation (iv) through the network results in two variable PNP identifications represented by additional green nodes. (b) Network-independent PNP identification relies on an efficient enumeration of all PNP variants (i) and further matching of spectra against these variants using the standard identification strategy (ii).

PNP identification strategies

The DEREPLICATOR algorithm¹¹ identified many PNPs in the GNPS dataset through standard identification (without modifications) and variable identification via spectral networks (Figure 1a). However, the limitation of the spectral network approach prevents DEREPLICATOR from finding many PNP variants. Indeed, variable identification via spectral networks works only when there exists an unmodified parent in a given connected component. Since PNPs vary across diverse related bacteria, this condition does not hold for many GNPS datasets because a PNP identified in one bacterium (and present in a database) is often represented by its modified variant in another bacterium. This limitation raises the challenge of developing methods for variable PNP identification that, in difference from DEREPLICATOR, do not rely on spectral networks.

Modification-tolerant search reveals diverse PNP variants

Since PNP databases are dominated by the most abundant representatives of PNP families, existing algorithms, focusing on identification of known PNPs, annotate only a small fraction of GNPS spectra. To address this limitation, we developed a network-independent VarQuest algorithm for modification-tolerant PNP identification (Figure 1b).

VarQuest revealed that a vast majority of PNP families (78%) identified in GNPS were not represented by any non-modified known PNP and thus are not detectable using the spectral network approach. This observation suggests that not only PNPs are extremely diverse across evolutionary distant microbial species but that also PNP families rapidly evolve so that PNP variants present in one species are often mutated/modified even in closely related species. This evolution of PNP families may reflect adaptation to unique ecological niches under various pressures, not unlike evolution of skyllamycins in Pseudomonas aeruginosa²⁰.

The great diversity of PNP variants underscores the importance of variable PNP identification via VarQuest and reveals a limitation of the spectral network approach implemented in DEREPLICATOR (most components in the GNPS spectral network turned out to be orphans). VarQuest has now fixed this unanticipated limitation. We benchmarked VarQuest and identified an order of magnitude more PNP variants as compared to existing PNP identification strategies.

Results

Brute-force approach

A novel PNP is called a variant of a known PNP if it has the same topology and sequence of amino acid, except for a single modified/mutated amino acid. We focus on identification of PNP variants with mass offset up to MaxMod (the default value MaxMod=300 Da).

The brute-force approach to variable identification (BruteForce) is based on enumeration of all possible modifications/ mutations for each peptide from the database²¹. Given a spectrum S and each PNP P from the database (with mass difference δ < MaxMod), it considers a modification of mass δ on all possible amino acids in P, forms a list CandidatePeptides(S) containing all such modified PNPs, and finds a PNP in CandidatePeptides(S) with the best match to S. Since the resulting list CandidatePeptides(S) contains a large portion of the entire PNP database, this approach is prohibitively time-consuming. Various database filtering strategies and spectral alignment algorithms were developed to speed-up the brute-force approach in traditional proteomics^{12–14, 22, 23}. However, extending variable identification algorithms from linear peptides to non-linear PNPs remains a challenge.

Spectral network approach

The spectral network approach (SpecNets) constructs the spectral network of all spectra and identifies the connected component of the spectral network that contains a given spectrum S (denoted Component(S)). It further forms the set CandidatePeptides(S) as the set of identifications of all spectra in Component(S) that were discovered using the standard identification method. Afterwards, it applies the spectral network propagation approach to CandidatePeptides(S) to perform variable identification of S. Although this approach is fast (since the CandidatePeptides(S) is typically small), it fails when Component(S) is an orphan, i.e., does not contain any spectra originating from known PNPs.

VarQuest algorithm

VarQuest pipeline for a single spectrum S (Figure 2) starts with selection of a short list CandidatePeptides(S) from the PNP database. Afterwards, VarQuest scores S against each PNP (with a single modification) in CandidatePeptides(S) and computes P-values of resulting PNP-spectrum matches (PSMs)²⁴. A peptide with the lowest P-value among all PNPs in CandidatePeptides(S) is reported as a candidate PNP that gave rise to the spectrum S.

Figure 2. VarQuest pipeline.

For a spectrum and a PNP database, VarQuest starts from scoring the spectrum against the entire database (i) to form the list of candidate PNPs. All possible modifications are considered for each candidate (ii) and the spectrum is scored against all variants (iii) to select the highest scoring variant per candidate PNP. Statistical significance of the scores is computed (iv) and the most statistically significant PSM is reported (v).

Efficient selection of the small list CandidatePeptides(S) is the key step of VarQuest. The standard identification approaches include a peptide P into CandidatePeptides(S) as long as Mass(S) ≈ Mass(P) with error up to Δ. Since Δ is small for high-resolution spectra, the list CandidatePeptides(S) is much smaller than the number of PNPs in the database, enabling fast standard identification but preventing detection of novel PNP variants. Var-Quest detects novel PNP variants by constructiong a short list CandidatePeptides(S) (in difference from a long list constructed by BruteForce) as described in the Methods section.

Although VarQuest searches for unknown PNPs with a single modification (as compared to a known PNP), it has ability to find PNPs with multiple modifications. However, in such cases, instead of reporting multiple modifications, it reports a single modification with combined mass equal to the total mass of multiple modifications. Below we illustrate how further analysis allows one to infer the positions and masses of multiple modifications.

Benchmarking VarQuest

We benchmarked VarQuest on five high-resolution GNPS datasets: Spectra_PSEUD (≈400,000 spectra from 260 Pseudomonas isolates²⁵), Spectra_STREP1 (≈200,000 spectra from Streptomyces⁷), Spectra_STREP2 (≈500,000 spectra from Streptomyces^{11, 26}), Spectra_CYANO (≈11 million spectra from Cyanobacteria²⁷), and Spectra_GNPS (≈130 million spectra from GNPS¹¹). Details on these datasets are provided in Supplementary Table 1.

We matched all spectral datasets against the PNPdatabase constructed by combining all PNPs from Anti-Marin²⁸, DNP²⁹, MIBiG³⁰, and StreptomeDB³¹ (5021 PNPs forming 1582 PNP families). The results were compared with BruteForce, SpecNets, and the standard identification (Standard) algorithms (Table 1). Time and memory requirements of these methods are described in Supplementary Table 2. Although BruteForce can find all VarQuest identifications (for a given P-value threshold) on the same set of spectra, it becomes prohibitively time-consuming even on moderately-sized spectral datasets such as Spectra_CYANO. Note that all methods are based on DEREPLICATOR¹¹ and that BruteForce and SpecNets failed to process the Spectra_CYANO and Spectra_GNPS datasets due to large memory requirements.

Table 1.

Comparison of four PNP identification approaches on various spectral datasets against the PNPdatabase with 5021 PNPs (1582 PNP families). Spectral networks were constructed using the GNPS interface (the Molecular Networking workflow²) with precursor and fragment ion tolerance set to 0.02 Da and all other parameters set to the default values. # PSMs stands for the number of identified PSMs, # variants stands for the number of identified PNP variants, and # PNPs (families) stands for the number of unique PNP identifications (PNP families). P⁻¹⁰, FDR₀ and FDR₅ stand for the number of identified PSMs, variants, unique PNPs or PNP families with P-value below 10⁻¹⁰, at 0%, and 5% False Discovery Rates, respectively. When processing Spectra_CYANO, the SpecNets tool crashed after a week of execution, and the BruteForce approach crashed due to exceeding memory limit (25 GB RAM).

Method	# PSMs			# variants			# PNPs (families)
	P−10	FDR0	FDR5	P−10	FDR0	FDR5	P−10	FDR0	FDR5
SpectraPSEUD
Standard	1029	997	1029	42	36	42	42 (13)	36 (12)	42 (13)
SpecNets	1837	1837	1837	164	164	164	41 (15)	41 (15)	41 (15)
BruteForce	7545	261	3822	1042	60	479	436 (197)	33 (13)	242 (117)
VarQuest	6055	224	6055	871	54	871	341 (148)	30 (13)	341 (148)
SpectraSTREP1
Standard	228	171	228	20	11	20	20 (10)	11 (5)	20 (10)
SpecNets	337	265	337	57	37	57	20 (10)	9 (3)	20 (10)
BruteForce	2648	57	841	1698	26	360	621 (293)	6 (1)	133 (69)
VarQuest	2266	128	716	1397	51	287	496 (225)	8 (3)	95 (49)
SpectraSTREP2
Standard	85	85	85	13	13	13	13 (5)	13 (5)	13 (5)
SpecNets	101	22	101	23	5	23	13 (5)	4 (2)	13 (5)
BruteForce	258	25	31	134	8	11	95 (60)	5 (2)	7 (3)
VarQuest	208	12	147	100	3	56	67 (45)	1 (1)	37 (25)
SpectraCYANO
Standard	353	146	353	42	23	42	42 (20)	23 (14)	42 (20)
VarQuest	3573	226	3573	2083	68	2083	702 (315)	25 (14)	702 (315)
SpectraGNPS
Standard	14757	7464	14746	420	279	420	420 (143)	279 (110)	420 (143)
VarQuest	379089	5661	179448	65204	1695	19619	2673 (835)	675 (256)	2025 (648)

We compared the number of identified PSMs and unique peptides for all methods at 0%and 5%False Discovery Rate (FDR) levels. To compute the FDR, VarQuest uses the concept of a decoy database extended to nonlinear peptides (see the Methods section). All PSMs with P-values above 10⁻¹⁰ were removed beforehand and the FDR was conservatively computed for the remaining PSMs. Table 1 shows more than ten times increase in the number of PSMs and five times increase in the number of PNPs and PNP families identified in GNPS via variable identification with VarQuest compared to the standard DEREPLICATOR at 5% FDR. Figure 3 shows the peptide network of the largest PNP family (cyclosporins) in the PNPdatabase identified by VarQuest (Supplementary Table 3).

Figure 3. Peptide network of the cyclosporin family in the PNPdatabase (a) extended by the newly identified cyclosporin variants (b).

The peptide network was constructed for 47 known cyclosporins using the GNPS interface (the Molecular Networking workflow²). Each node represents theoretical spectrum of a cyclosporin variant, the number inside each node stands for monoisotopic mass in Da rounded to integers. Two nodes are connected by an edge if the corresponding theoretical spectra are similar (cosine score at least 0.8). Blue edges corresponds to characteristic mass shifts of 14, 16, 28, 32, and 42 Da, the remaining edges are black. (a) Peptide network of cyclosporin variants present in the PNPdatabase. Red nodes are 29 cyclosporins identified in Spectra_GNPS as known PNPs (both by DEREPLICATOR and VarQuest). Blue nodes are theoretical spectra of the rest 18 PNPs which are not present in GNPS in their known form and added to the network for the sake of completeness. (b) Peptide network of known and novel cyclosporin variants. Green nodes are theoretical spectra of 18 novel variants identified by VarQuest in GNPS. Each novel variant is the most statistically significant identification of the corresponding blue node (an absent known cyclosporin PNP). Likely insertions/deletions are shown on corresponding edges, Hiv stands for Hydroxyisovaleric acid, mLeu stands for methylated leucine.

While DEREPLICATOR revealed spectra corresponding to only 8% of peptides in the PNPdatabase, VarQuest increased this number to 40%. 1605 out of 2025 PNPs identified by VarQuest in Spectra_GNPS (at 5% FDR) have their unknown variants present in Spectra_GNPS, while their known variants are absent and thus can not be detected by the standard identification strategies.

Our analysis of the entire GNPS dataset at 5% FDR identified 648 PNP families (41% of all known PNP families). At the same time, only a small fraction of identified PNP families (143 out of 648) were identified as an unmodified parent, i.e., a vast majority of identified PNP families (78%) were not represented by any non-modified peptides in the PNPdatabase and thus are not detectable using the spectral network approach.

Modifications/mutations identified by VarQuest

Supplementary Table 4 shows the most common mass offsets identified by VarQuest in the Spectra_GNPS dataset at 5% FDR. For each mass offset we identified its most likely position in the PNP. As expected, the most common offsets are −14 Da (demethylation), +14 Da (methylation), +28 Da (dimethylation), +18 Da (hydration), and +16 Da (hydroxylation) corresponding to either modifications/adducts or mutations. In addition, Supplementary Table 4 reveals many surprising offsets such as −95 Da offset (primarily at Leucine/Isoleucine), −81 Da offset (primarily at Valine), and others. These offsets may correspond to the combination of the amino acid loss (−113 Da for Leucine/Isoleucine and −99 Da for Valine) and hydration (+18 Da). The abundance of such offsets suggests that the recently described phenomenon of amino acid deletions/insertions in NRPs²⁵ (due to A-domain stuttering and skipping in NRPSs³²) may be more prevalent than previously thought. Genome mining efforts typically rule out such events due to the consecutive arrangements of A-domains in NRP synthetases.

To conservatively estimate the number of indels revealed by spectra in Spectra_GNPS, we considered identified mass offsets that matched monoisotopic masses of proteinogenic amino acids within a 0.02 Da error. This analysis revealed 217 (169) putative insertions (deletions) out of 19619 PNP variants identified by VarQuest in Spectra_GNPS at 5% FDR. Our confidence in the deletions is higher than in the insertions since for each of them we also checked that the deleted amino acid is present in the known PNP structure (which reduced the initial number of potential deletions by 30% from 242 to 169).

Analysis of PNP diversity

VarQuest identified 19619 PNP variants related to 2025 distinct PNPs in Spectra_GNPS at 5% FDR. More than 70% of the identified PNPs (1489) were found in at least two different forms. Each identified PNP was found in 9.7 various PNP forms on average with the maximum value equal to 239 for tolybyssidin A. Our analysis adds a “chemical” dimension to the recently revealed PNP diversity at the biosynthetic gene cluster (BGC) level^{33, 34}. We further revealed that related bacteria are likely to produce similar PNP variants rather than identical PNPs (see the Methods section).

Validation of VarQuest identifications

Our analysis revealed that about 60% of the PNP variants identified by VarQuest in Pseudomonas and Streptomyces datasets are missed by DEREPLICATOR and SpecNets. We validated the most statistically significant VarQuest hits using literature search for identified PNP origin which should correlate with the sample origin (see the Methods section) and searched for BGCs by genome mining³⁵ whenever the genome of the analyzed species is available. We further analyzed three identified PNP variants (referred to as Massetolide-1252, Venepeptide-2154, and Surugamide-769) in more details (Supplementary Figure 1).

Massetolide-1252

Massetolide A is a known NRP from Pseudomonas³⁶ that consists of a cycle TISLSLI and a branch EL (along with 3-hydroxydecanoic acid lipid tail of mass 171 Da) attached to the cycle via a bond connecting T in the cycle with E in the branch. We represent branch-cyclic peptides as a concatenate of its cyclic sequence and its branch sequence, both starting from their attachment points, e.g., massetolide A is represented as TISLSLI*EL. VarQuest identified massetolide A and its novel variant Massetolide-1252 (sequence TISL⁺¹¹³SLI*EL and mass 1252.8 Da) with P-value 4.2 · 10⁻¹⁹ using a spectrum from P. synxantha. The +113 Da offset corresponds to insertion of Leucine or Isoleucine residue and matches the recently identified poaeamide B with sequence TISLLSLI*EL and mass 1252.8 Da. Note that a single run of VarQuest instantly achieved the same goal as the time-consuming semi-manual discovery of poaeamide B²⁵.

VarQuest also rediscovered bananamides, a family of PNPs discovered in the same study²⁵. Bananamide (referred to as Bananamide-1093) was identified with P-value 4.3 · 10⁻¹⁰ as a variant of massetolide A (sequence TIS⁻⁴⁶LSLI*EL and mass 1093.7 Da) using a spectrum from P. fluorescens. While the recent study²⁵ did not derive the amino acid sequence from this spectrum, it purified and sequenced a related PNP (named bananamide 2) with sequence TLLQLI*DL (along with C12 3-hydroxy unsaturated acid lipid tail of mass 197 Da) and mass 1105.7 Da (amino acids differing from massetolide A are highlighted except for a change between amino acids I and L with identical masses). While amino acid sequences TIS⁻⁴⁶LSLI*EL and TLLQLI*DL appear to be rather different, note that S⁻⁴⁶LS has the same mass as LQ suggesting that TIS⁻⁴⁶LSLI*EL may actually correspond to TILQLI*EL with a single deleted amino acid as compared to massetolide A. Note that there is only one difference with respect to masses of amino acids between this sequence (TILQLI*EL) and the sequence of bananamide 2 (TLLQLI*DL).

Our analysis of Bananamide-1093 suggests that bananamides emerged from the massetolides family after deletion of a single amino acid (or alternatively, massetolides emerged from bananamides after insertion of a single amino acid). Interestingly, while the PSM for Bananamide-1093 is statistically significant, PSMs for bananamides 1, 2, and 3 identified in²⁵ have rather high P-values that did not pass the VarQuest P-value threshold. Remarkably, the manual analysis in²⁵ missed the most statistically significant PSM for bananamides identified by VarQuest, illustrating the power of automated approaches to PNP identification. Moreover, after identifying Bananamide-1093, VarQuest identifies spectra of bananamides 1, 2, and 3 against Bananamide-1093 as statistically significant PSMs with P-values 1.1 · 10⁻¹³, 9.8 · 10⁻¹², and 1.1 · 10⁻¹⁶, respectively.

Surugamide-769

Surugamides are cyclic NRPs from marine streptomyces^{11, 37}. VarQuest identified both known PNP surugamide B with sequence IAIVKIFL and its novel variant IAIVK⁻¹²⁸IFL using a spectrum from S. albus (P-value 1.7 · 10⁻¹⁹). The SpecNets approach missed this compound because its connected component does not contain known surugamides.

The amino acid sequence IAIVK⁻¹²⁸IFL of Surugamide-769 corresponds to a loss of Lysine. This annotation is consistent with the arrangement of the genes in the surugamide BGC since the deleted Lysine corresponds to the last A-domain in one of two genes in this BGC (Supplementary Figure 2). Thus, Surugamide-769 represents the second evidence of the same NRP synthetase producing two cyclic peptides with different numbers of amino acids, similar to poaeamide B and massetolide A²⁵. However, further experimental validation of this hypothesis and many other likely insertions and deletions listed in Supplementary Table 4 is needed.

Venepeptide-2154

Venepeptide is a linear ribosomal peptide M⁺²⁸NVITNLLAGVVHFLGWLV that was identified from S. venezuelae³⁸. VarQuest identified its variantM⁺²⁸NVITN⁺³¹LLAGVVHFLGWLV (mass 2154.1 Da) with P-value 3.2 · 10⁻¹⁵ using a spectrum from S. lividans. DEREPLICATOR missed this compound because GNPS does not contain a spectrum corresponding to the known venepeptide. Sequence similarity search³⁹ of this peptide against the genome of S. lividans revealed the sequence MNLLTDILAGLVHFVGWLV (the differences with venepeptide are highlighted). A match of the spectrum against this sequence resulted in a PSM with P-value 8.5 · 10⁻²⁴ and suggested modification +44 Da on the M residue. Note that while VarQuest is limited to finding variants with a single modified amino acid, it was able to identify that a spectrum from S. lividans has arisen from a variant of venepeptide. The further manual analysis revealed that Venepeptide-2154 structure differs from venepeptide in four amino acids.

Discussion

Although the launch of high-throughput natural products discovery pipelines, such as the GNPS molecular network, is an important step towards future discoveries, the lack of computational approaches is still a bottleneck for spectral identifications in the GNPS infrastructure. Currently, the GNPS spectral library, a collection of identified spectra from GNPS, represents a minuscule fraction of all GNPS spectra. While molecular networks^{2, 19} have already resulted in discoveries of various PNPs and their variants^{25, 40}, these discoveries still requires time-consuming manual follow-up analysis. Here we demonstrated how the same goal can be achieved in a single push-of-a-button VarQuest run, replicating recent PNP discoveries and finding previously unknown PNP variants. Moreover, variable dereplication of the entire GNPS revealed both surprising diversity of PNPs and limitations of the spectral networks approach. In particular, we demonstrated that the recently discovered phenomenon of insertions and deletions of amino acids is widespread among NRPs.

There is a yet another reason why variable identification is important. Recent Genomic Encyclopedia of Bacteria and Archaea study⁴¹ revealed many BGCs with PNPs representing the largest group of secondary metabolites encoded by these BGCs. However, the vast majority of the predicted BGC products remained unknown, reflecting the limited information available for characterized natural products and the lack of genome mining and peptidogenomics tools for matching BGCs and spectra.

While databases in traditional proteomics consist of known peptides, the ongoing genome mining efforts for PNP discovery³⁵ generate vast databases of still uncharacterized putative PNPs^{7, 42, 43}. Since predicting an NRP encoded by an NRPS is a difficult problem, various tools for predicting specificities of A-domains⁴⁴ output multiple rather than a single candidate amino acid for each A-domain. Supplementary Figure 2 presents three top candidate amino acids for each of eight A-domains in suragamide-encoding NRPS resulting in 3⁸ candidate NRPs. As the result, genome mining efforts typically generate large databases of error-prone putative PNPs, and matching spectra against such databases is prohibitively time-consuming. Thus, development of fast algorithms for variable PNP identification is important for the success of genome mining efforts.

We have presented VarQuest algorithm for variable PNP identification via database search of mass spectra, the only modification-tolerant approach capable of searching the entire GNPS spectral network. Our method revealed an order of magnitude more PNPs than the standard search by DEREPLICATOR illuminating the “dark matter of PNPs”⁴⁵. It also greatly increased the spectral library of PNPs in GNPS by identifying 41% of all known PNP families in the PNPdatabase. Iterative run of VarQuest has a potential to identify even more PNP variants with multiple modifications.

VarQuest revealed a surprising diversity of PNPs that may reflect evolutionary adaptation of various bacterial species to changing environment and competition, e.g., a continuous change of the repertoire of variants of peptidic antibiotics in response to developing antibiotic resistance. It also revealed a limitation of existing NRP mining tools that were developed based on “NRPS – a single NRP” pairs as the training datasets⁴⁴ aimed at predicting a single NRP. A more biologically adequate approach would be to use the training datasets “NRPS – NRP network” that have recently become available. With growing availability of paired genomics and mass spectrometry datasets, it is now possible to generate such training datasets using VarQuest.

Methods

Scoring PNP-spectrum matches

A PNP graph of a PNP P is defined as a graph with nodes corresponding to amino acids in P and edges corresponding to generalized peptide bonds¹¹. The mass of a PNP graph (referred to as mass(P)) is defined as the total mass of its amino acids and TheoreticalSpectrum(P) is defined as the set of masses (theoretical peaks) of all connected components of the PNP graph resulting from removal of two edges (a 2-cut in cyclic and branch-cyclic PNPs) or a single edge (a bridge in a branch-cyclic PNP)¹¹. Note that each such removal results in two peaks with total mass equal to Mass(P).

Given a peptide P and a spectrum S, SPCScore(P, S) is defined as the Shared Peak Count, the number of peaks shared between TheoreticalSpectrum(P) and S. Two peaks are shared if their masses are within a threshold ε (0.02 Da for high-resolution spectra). We compute this score only if the precursor mass of the spectrum, denoted as Mass(S), matches Mass(P) with error up to Δ (0.02 Da for high-resolution spectra).

If (A₁, . . . , A_n) is the list of amino acid masses in a PNP P, we define Variant (P, i, δ) as (A₁, . . . , A_i + δ, . . . , A_n), where P and Variant (P, i, δ) have the same topology and A_i + δ ≥ 0. VariableScore(P, S) is defined as

$$\mathit{\text{max}}(\mathit{\text{SPCScore}}(\mathit{\text{Variant}}(P,i,\omega ),S)),$$ (1)

where ω is Mass(P) − Mass(S) and i varies from 1 to |P| (|P| stands for the number of amino acids in the peptide P). We define a variant of peptide P derived from a spectrum S (referred to as Variant(P, S)) as Variant(P, i, ω) of peptide P that maximizes SPCScore(Variant(P, i, ω), S) across all positions i in P.

Selecting candidate peptides

Consider a peptide P and its variant P^* = Variant(P, i, δ). TheoreticalSpectrum(P) and TheoreticalSpectrum(P^*) share approximately half of their peaks while the remaining peaks in TheoreticalSpectrum(P^* are shifted by δ with respect to the corresponding peaks in TheoreticalSpectrum(P) (Supplementary Figure 3). Thus, if an experimental spectrum S is produced by a peptide P^* and shares N peaks with TheoreticalSpectrum(P^*), we expect that $\mathit{\text{SPCScore}}(P,S)\approx \frac{N}{2}$. However, this condition often does not hold in practice due to many noisy and missing peaks in experimental spectra. In practice, we reduce the size of CandidatePeptides(S) by retaining all PNPs that satisfy the condition:

$$\mathit{\text{SPCScore}}(P,S)\ge \eta ,$$ (2)

for a small value η. To select the threshold η, we analyzed the values of SPCScore for peptides reported by the brute-force method at various significance levels (Supplementary Table 5). Since the vast majority of statistically significant PSMs (P-value ≤ 10⁻¹⁰) share at least 5 peaks with the corresponding known peptides (74%, 80% and 72% for Spectra_PSEUD, Spectra_STREP1 and Spectra_STREP2, respectively), we set the default value η = 5.

For a given spectrum S, VarQuest forms the list CandidatePeptides(S) by selecting all PNPs satisfying the equation 2 (among all PNPs with mass differing from Mass(S) by at most MaxMod). Checking this condition requires computing SPCScore(P, S) values for each peptide P from the PNP database Peptides. Since a naive approach (computing SPCScore for each PNP) is time consuming, VarQuest preprocesses the PNP database, and scores a spectrum S against the entire database at once.

Preprocessing a PNP database

For a given PNP database Peptides, the preprocessing starts from generation of theoretical spectra for each PNP in the database (Stage 1 in Supplementary Figure 4). All peaks from all theoretical specta are combined altogether and sorted to form the array SortedPeaks(Peptides) (Stage 2). The peaks in SortedPeaks(Peptides) are partitioned into M bins of size θ (the default values M = 20000 and 8 = 0.2 Da). Afterwards, VarQuest constructs an indexing table Index(Peptides, M, θ) (Stage 3). The table is designed in such way that the i-th cell Index[i] contains a pointer to the smallest peak p, such that p ≥ i · θ for all i ∈ [0..M − 1].

Scoring a spectrum against a PNP database

To score a given spectrum S against all PNPs in a PNP database Peptides, VarQuest iterates through all the peaks in S and stores PNPs matching the peak into a counting set FeasiblePeptides(S) (Supplementary Figure 5). To match a peak s against all the PNPs, VarQuest counts all the matching theoretical peaks in the interval (s − ε, s + ε) by finding p_lower (the smallest matched peak) and p_upper (the largest matched peak). VarQuest sets $i_{\mathit{\text{lower}}}=\lfloor \frac{s-\epsilon }{\theta }\rfloor$ and uses binary search to search for the smallest matching peak in the interval between Index[i_lower] and Index[i_lower + 1] (p_upper is found in a similar way). Since the interval is small, the binary search is much faster than the search on the entire array SortedPeaks(Peptides).

After processing all peaks in the spectrum S, the number of occurrences of a PNP P in FeasiblePeptides(S) corresponds to the number of shared peaks between P and S. Thus, the list FeasiblePeptides(S) contains information about SPCScore(P, S) for all PNPs P sharing at least one theoretical peak with S.

Computing false discovery rate

The target-decoy approach⁴⁶ for estimating FDR is based on generating a decoy database DecoyPeptides from a target database Peptides and searching all spectra against combined DecoyPeptides and Peptides databases. The target-decoy approach further uses the numbers of PSMs found in both databases to evaluate FDR. We refer to the set of all PSMs found in Peptides (DecoyPeptides) and having P-values below τ as PSM_τ(Peptides, Spectra) (PSM_τ(DecoyPeptides, Spectra)). As the decoy database consists of randomly generated peptides, we expect to find very few PSMs in PSM_τ(DecoyPeptides, Spectra) for an appropriately chosen τ. Note that the size of DecoyPeptides is not necessary equal to the size of Peptides. We consider the situation when the frequencies of target and decoy peptides in the combined database are t and d, respectively (t + d = 1). We define the decoy ratio D as $\frac{d}{t}$ and compute FDR as follows:

$${\mathit{\text{FDR}}}_{\tau }=\frac{1}{D}\frac{|{\mathit{\text{PSM}}}_{\tau }(\mathit{\text{DecoyPeptides}},\mathit{\text{Spectra}})|}{|{\mathit{\text{PSM}}}_{\tau }(\mathit{\text{Peptides}},\mathit{\text{Spectra}})|}.$$ (3)

Since VarQuest algorithm is linear with respect to the size of the PNP database, larger DecoyPeptides lead to increased running time. On the other hand, small database DecoyPeptides may result in an inaccurate estimate of FDR. We thus benchmarked VarQuest with various values of D to show that D = 1 is a good trade-off (Supplementary Table 6).

Generating decoy database

A popular method for generating decoy databases in traditional proteomics is random shuffling of amino acids for each target protein. However this strategy (Supplementary Figure 6b, further referred as Classical) is not suitable for PNPs because (i) PNPs are much smaller than proteins, (ii) many PNPs are cyclic or branch-cyclic, and (iii) many PNPs contain multiple copies of the same amino acid (Supplementary Figure 7). This results in decoy peptides that are similar to the target peptides after the shuffling procedure, resulting in an inflated FDR.

To address this challenge, DEREPLICATOR¹¹ randomly redistributes the total mass of a peptide over the nodes of its PNP graph (Supplementary Figure 6c, DEREPLICATOR strategy). This strategy is motivated by the fact that PNPs often contain non-standard amino acids with a wide range of masses.

VarQuest uses a novel decoy generation approach based on amino acid shuffling and random bond displacement (Supplementary Figure 6d, VarQuest strategy). For each target PNP, VarQuest first generates a decoy PNP by rearranging amino acids. Afterwards, it randomly selects an edge in the PNP graph and substitutes it by a new edge, connected to a randomly selected position, such that the resulting decoy structure represents a connected graph. This strategy takes into account the complex structures present in many PNPs, resulting in a more diverse decoy database.

To compare accuracy of the FDR estimation using these methods we conducted the following experiment. We took 200 top-scoring unique PNP identifications from DEREPLICATOR run on the entire GNPS¹¹. These annotations were manually curated and validated as reliable. In Experiment 1, we ran VarQuest on the spectra with the same PNP database as in¹¹. In Experiment 2, we excluded 200 target peptides and all their known variants from the PNP database and ran the VarQuest again. We expect FDR around 0% in Experiment 1 (all mass spectra are highly trustable) and around 50% in Experiment 2 (the correct peptides are missing from the database, and matches to the target and decoy PNPs are equally likely). Supplementary Table 7 shows FDR estimations for both experiments computed based on various decoy generation approaches. Classical strategy overestimates FDR in Experiment 1 while DEREPLICATOR method underestimates FDR in Experiment 2. VarQuest decoy generation strategy has an acceptable performance in both cases (0.5% and 55.0% respectively).

Constructing the PNPdatabase

We combined all compounds with at least 4 generalized peptide bonds from AntiMarin²⁸, DNP²⁹, MIBiG³⁰, and StreptomeDB³¹ into a single non-redundant database with 10067 distinct compounds (Supplementary Table 8). These chemical entities were classified into chemical classes using ClassyFire⁴⁷ software tool. Compounds related to peptidic classes were included into our target database (referred to as PN-Pdatabase) that consists of 5021 distinct PNPs forming 1582 PNP families (Supplementary Table 9). Supplementary Tables 10–14 show distributions of PNP origins, PNP family sizes, PNP structures, number of peptide bonds and the most frequent amino acids in the PNPdatabase.

Revealing PNP diversity in related bacteria

Spectral libraries in metabolomics^{2, 48} rely on the comparative metabolomics assumption that assumes that two related bacteria are likely to produce identical metabolites. Our analysis revealed that, in the case of PNPs, such cases are relatively rare and that related bacteria are likely to produce similar rather than identical PNPs. To illustrate this point, we visualized strain relations based on DEREPLICATOR (identical known PNPs) and VarQuest (PNP variants of the same origin) identifications in Spectra_CYANO (the largest) and Spectra_STREP1 (the less contaminated) datasets at 5% FDR. To illustrate the diversity of PNPs across related bacteria, we introduced the concept of the strain graph with nodes representing strains and edges connecting two strains if they produce variants of the same known PNP (see Supplementary Figure 8)

Spectra_CYANO

DEREPLICATOR identified 42 known PNPs in 68 out of 352 Cyanobacteria strains. The strain graph constructed on these PNPs has 284 edges (two strains share identical PNP) and consist of 13 connected components (Supplementary Figure 8a). VarQuest detected PNP variants of 334 known PNPs in the same set of 68 strains. The strain graph has 618 edges (two strains produce PNP variants of the same known PNP) and a single connected component (Supplementary Figure 8b). VarQuest strain graph on the entire Spectra_CYANO contains 272 nodes, 3791 edges and 19 connected components.

Spectra_STREP1

DEREPLICATOR identified 20 PNPs in 10 out of 17 Streptomyces strains. Its strain graph has only 6 edges and consists of 6 connected components (Supplementary Figure 8c). In contrast, VarQuest identified 78 PNP variants in these 10 strains and enlarged the graph by 29 additional edges (35 total) turning it into a single connected component (Supplementary Figure 8d). Moreover, VarQuest was able to identify PNP variants in all 17 Streptomyces strains in this dataset. The full strain graph has 73 edges and 3 connected components.

Validating VarQuest identifications using literature search

Supplementary Table 15 shows the list of 244 peptide variants identified by VarQuest in Pseudomonas and Streptomyces datasets. We considered all PNP variants at 5% FDR (871, 287, and 56 for Spectra_PSEUD, Spectra_STREP1, and Spectra_STREP2, respectively), and excluded identifications of known PNPs with zero mass offsets (resulting in 662, 239, and 43 remaining peptide variants, respectively). Afterwards, we analyzed 100 peptide variants with the lowest P-values per dataset (for Spectra_STREP2 we considered all 43 variants). Origin of each PNP family was determined based on literature search. The most contaminated dataset is Spectra_PSEUD where only 52% of variants have Pseudomonas origin. Four large non-Pseudomonas families (Surfactins, Xentrivalpeptides, Bacillomycins and SNA-60-367) cover 26 variants in this dataset. SpecNets identified 19 out of these 26 variants which indirectly suggests that these spectra are true contaminants rather than VarQuest false positives. Half of the singleton (PNP families with a single identified member of the family) contaminants (10 out of 22) are also reported by SpecNets. Both Streptomyces datasets have higher rate of PNPs originally found in Streptomyces (94% for Spectra_STREP1 and 65% for Spectra_STREP2).

There are a few reasons why spectra from Spectra_PSEUD dataset form PSMs with PNPs from other bacterial sources apart from being false PSMs, e.g., laboratory contamination and morphology misidentification as many laboratory collections contain organisms that are misidentified¹¹. Also, Luria Broth growth media prior to autoclaving is not sterile, e.g., surfactins are commonly found in the growth media (even freshly opened bottles).

Running VarQuest iteratively

While VarQuest is limited to searching for PNP variants with a single modification, this limitation can be potentially addressed by an iterative run of VarQuest. In this case, PNP variants identified in the initial VarQuest run are iteratively used as an input PNP database for the subsequent run on the same spectral dataset. Supplementary Table 16 presents results of iterative VarQuest run on Spectra_CYANO. On the initial run on this dataset, VarQuest identified 2083 and 95 PNP variants in the target and decoy versions of the PNPdatabase, respectively. For the second iteration, we selected PNP variants with the most reliable mass offsets equal to ±14 Da (methylation), ±28 Da (dimethylation), ±18 Da (hydration), and ±16 Da (hydroxylation) and ended up with a new PNP database with 81 PNP variants (78 targets and 3 decoys) representing 53 unique PNPs. We further refer to this database as FirstIterationDB.

Out of 385 PNP variants identified on the second iteration, 69 were already identified among 2083 PNP variants reported in the first VarQuest run and the remaining 284 are novel (14% increase). We investigated why 353 PNP variants identified at the 2nd iteration were not reported on the first run of VarQuest (Supplementary Figure 9). It turned out that a large fraction of newly identified PNP variants (114 out of 353) were actually identified (but not reported) by VarQuest since they have P-values slightly above the default P-value threshold of 10⁻¹⁰ (varying from 10⁻¹⁰ to 10⁻⁷). Thus, FirstIterationDB indeed presents a better PNP database for identifying PNPs with multiple modifications as compared to the original PNP database.

To provide additional evidence that the newly found PNP variants represent correct rather than erroneous modifications, we further checked if the most frequent offsets identified in the 2nd iteration are consistent with the most frequent offsets identified in the initial run of VarQuest (Supplementary Table 17). It turned out that most of these offsets correlate with the most common offsets identified by VarQuest in Spectra_GNPS (Supplementary Table 4).

Code availability

VarQuest is available both as a command line tool (http://cab.spbu.ru/software/varquest) and as a web application at the GNPS website (http://gnps.ucsd.edu).

Data availability

LC–MS/MS data are publicly accessible under MassIVE accession nos. MSV000079450 (Spectra_PSEUD), MSV000078604 (Spectra_STREP1), MSV000078839 (Spectra_STREP2), and MSV000078568 (Spectra_CYANO) at http://gnps.ucsd.edu/ProteoSAFe/datasets.jsp. The list of 120 MassIVE accessions numbrers for Spectra_GNPS is available at http://cab.spbu.ru/software/varquest. The PNPdatabase is available at http://cab.spbu.ru/software/varquest.